Boundary detection in petrographic images and applications of S-transform space-wavenumber analysis to image processing for texture definition.

摘要

A petrographic thin section is a 30 micron thick slice of a rock mounted on a glass microscope slide, usually composed of randomly oriented crystalline, mineral grains. Viewed in cross-polarized light, the mineral grains exhibit characteristic colors and textures. When the polarizer and the analyzer are rotated, each mineral grain undergoes separate and independent cycles of brightness and extinction. The colors and textures are used by geologists to identify the grains and the mineral composition of a rock. This thesis attempts to automate the process of grain recognition through use of digital imaging, numerical algorithms, and a proposed texture characterization procedure using the local S-transform time-frequency (or space-wavenumber) spectrum.; Conventional methods of grain recognition are based on several separate images, which are obtained by rotating the thin section between crossed polarizers. In this thesis, a novel approach to identifying grains is proposed; it reduces the input images to one color image plus one gray-level image. The two images are synthesized by mapping the maximum color intensity (max-image) and the corresponding angular rotation, represented by a gray level (phi-image). Two modified Canny edge operators are used to detect edges in the max- and phi-images separately. A seeded region growing algorithm is developed to find boundaries in the max- and phi-images based on the edge information. The two boundary maps are finally combined into a single boundary map in which each region corresponds to a grain. Three sets of petrographic images are used to test the method.; A texture is usually characterized by several dominant frequencies within a certain bandwidth. Thus, texture segmentation can be completely based on local Fourier spectra. The polar S-transform, developed in this thesis, provides a multi-resolution and rotation-invariant local spectrum. A framework for numeric computation of the polar S-transform that allows the polar S-transform to be easily implemented on most computers is also proposed. (Abstract shortened by UMI.)